Acta Materialia 166 (2019) 335e345

Contents lists available at ScienceDirect

Acta Materialia
journal homepage: www.elsevier.com/locate/actamat

Full length article

Establishing structure-property localization linkages for elastic

deformation of three-dimensional high contrast composites using
deep learning approaches
Zijiang Yang a, Yuksel C. Yabansu b, Dipendra Jha a, Wei-keng Liao a, Alok N. Choudhary a,
Surya R. Kalidindi b, c, Ankit Agrawal a, *
a
Department of Electrical Engineering and Computer Science, Northwestern University, Evanston, IL, 60208, USA
b
George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, 30332, USA
c
School of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, GA, 30332, USA

a r t i c l e i n f o a b s t r a c t

Article history: Data-driven methods are attracting growing attention in the field of materials science. In particular, it is
Received 24 July 2018 now becoming clear that machine learning approaches offer a unique avenue for successfully mining
Received in revised form practically useful process-structure-property (PSP) linkages from a variety of materials data. Most pre-
19 December 2018
vious efforts in this direction have relied on feature design (i.e., the identification of the salient features
Accepted 22 December 2018
Available online 31 December 2018
of the material microstructure to be included in the PSP linkages). However due to the rich complexity of
features in most heterogeneous materials systems, it has been difficult to identify a set of consistent
features that are transferable from one material system to another. With flexible architecture and
Keywords:
Materials informatics
remarkable learning capability, the emergent deep learning approaches offer a new path forward that
Convolutional neural networks circumvents the feature design step. In this work, we demonstrate the implementation of a deep learning
Deep learning feature-engineering-free approach to the prediction of the microscale elastic strain field in a given three-
Localization dimensional voxel-based microstructure of a high-contrast two-phase composite. The results show that
Structure-property linkages deep learning approaches can implicitly learn salient information about local neighborhood details, and
significantly outperform state-of-the-art methods.
© 2019 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

1. Introduction scale. Most often this involves capturing the effective (i.e., ho-
mogenized) properties defined at the higher material structure
Material systems used in advanced technologies often exhibit scale, while accounting for the heterogeneity that exists at the
rich heterogeneity over a hierarchy of well-separated length scales. lower material structure scale. A large number of strategies have
Most material properties are influenced strongly by certain, not yet been developed and established in literature for addressing ho-
clearly identified, details of this heterogeneity in the material mogenization, ranging from the elementary bounding theories
structure. However, the natural hierarchy of the well-separated [5e7] to the self-consistent approaches [8e10] to the sophisticated
material structure length scales allows the adoption of hierarchi- statistical continuum theories [11e13].
cal multiscale modeling approaches [1e4]. Central to these ap- Localization, on the other hand, deals with the transfer of salient
proaches is the efficient communication of the salient information information from a higher material structure scale to a lower ma-
between the hierarchical structure scales with a sharp focus on the terial structure scale. As a specific example, it might deal with the
details that strongly influence the overall properties and perfor- microscale spatial distribution of a macroscopically imposed stress
mance characteristics of the material. (or strain rate) tensor. An illustration of localization in hierarchical
Homogenization deals with the transfer of salient information multiscale modeling is shown in Fig. 1. As such, localization prob-
from a lower material structure scale to a higher material structure lems are much more difficult compared to homogenization, and
have received only limited attention in current literature [14e23].
Indeed, localization plays an important role in the assessment or
* Corresponding author. prediction of failure-related properties. For instance, both high-
E-mail address: [email protected] (A. Agrawal). and low-cycle fatigue properties of metals are strongly affected by

https://doi.org/10.1016/j.actamat.2018.12.045
1359-6454/© 2019 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
336 Z. Yang et al. / Acta Materialia 166 (2019) 335e345

Fig. 1. Illustration of localization in hierarchical multiscale modeling.

the localization of the stress and plastic strain at the microscale property linkages [42e45], microstructure reconstructions
[24e26]. Since the underlying physics behind both homogenization [46,47], and microstructure generation [48,49] with generative
and localization problems is essentially described by exactly the adversarial networks (GAN) [50]. In contrast to traditional machine
same set of governing field equations, it stands to reason that the learning approaches, deep learning offers an end-to-end frame-
localization solutions often automatically embed homogenization work where it usually takes raw data as input (i.e., without the need
solutions within them [27e29]. for any feature engineering). Because of its flexible structure and
Localization problems have largely been addressed in literature remarkable learning capability, deep learning can automatically
using either numerical approaches (e.g., the finite element method extract higher-order information embedded in the raw input. It was
(FEM) [30,31]) or iterative methods employing Green's functions shown in prior work [44] that convolutional neural networks (CNN)
and fast Fourier transforms (FFTs) [14,16,32]. Neither of these ap- can automatically extract important higher-order microstructure
proaches are particularly suited to fast exploration of the vast statistics (i.e., up to 1000-point statistics) central to establishing
microstructure design space because of their relatively high reliable homogenization structure-property linkages for high
computational cost. One of the main drawbacks of these ap- contrast material systems. The extraction of this level of informa-
proaches is that they do not focus on learning transferable tion is not computationally feasible through standard feature
knowledge from one microstructure to another microstructure. In extraction methods. In this work, we focus on two high-contrast
other words, they are not formulated to take advantage of previ- composite material systems, with contrast values of 10 and 50
ously aggregated information on different microstructures to make (these represent the ratios of Young's moduli of the constituent
predictions for new microstructures. One of the most promising phases in the composite). More specifically, we propose a deep
approaches to the localization problem is the Materials Knowledge learning model to efficiently and accurately predict microscale
Systems (MKS) approach [17e20,22e24,33] that employs cali- elastic strain field of three-dimensional (3-D) high contrast elastic
brated Green's function based kernels in a non-iterative series so- composites. In order to critically evaluate the performance of pro-
lution. The main challenge comes from the difficulty in engineering posed model, the results obtained via deep learning approach are
the local features needed to establish highly accurate localization compared to three different benchmark approaches that solve the
relationships. In general, the features of interest in establishing localization problem via other data-driven techniques [19,20,22].
localization relationships are all related to the local neighborhood
details at the focal voxel of interest (i.e., the voxel under exami- 2. Benchmark methods for localization
nation) in a given microstructure. In high-contrast composites, the
local interactions between microscale voxels extend to larger Elastic localization has been modeled for 3-D microstructures
lengths and demand the consideration of larger neighborhood re- through single and multi-agent feature extraction methods [19,20],
gions. It was shown in prior work [23] that the number of features and the materials knowledge system (MKS) framework [22]. These
that need to be defined for the larger microstructure neighborhood methods are used in this study as benchmarks to evaluate the
explodes exponentially, posing significant hurdles in the calibration performance of deep learning methods, and are briefly described
of the kernels through standard machine learning approaches. next.
In recent decades, deep learning approaches have proven their
superior performance over traditional machine learning ap-
2.1. Feature extraction methods
proaches [34e36]. Deep learning techniques have also recently
gained attention in the materials science field, and have been
The first two benchmarks used in this paper are based on feature
successfully employed in materials image segmentation [37,38],
extraction methods [19,20]. Liu et al. [19] extracted a relatively
materials data prediction [39e41], homogenization structure-
large set of features to identify the local neighborhood around the
 Z. Yang et al. / Acta Materialia 166 (2019) 335e345 337

focal voxel of interest, applied feature ranking methods to establish various difficult problems related to feature recognition and image
a reduced set of features that preserve the most significant infor- segmentation [57e62]. A conventional CNN model usually has a
mation, and then used random forest techniques to predict stack of several convolutional layers and pooling layers. The con-
microscale elastic strain field. However, such a single-agent feature volutional layers employ filters to learn the important local features
extraction method does not effectively capture the embedded of images along with their relative importance. The pooling layers
correlations in the trained data. Thus, in a following study [20], Liu reduce dimensionality, while preserving the most important in-
et al. developed a multi-agent feature extraction method to further formation. The output of multiple stacks of convolutional and
improve the model's performance. More specifically, this method pooling layers are then fed into fully connected layers which have
consisted of two steps. In the first step, the ensemble of micro- the final list of predictors on which the model is trained. In this
structures were divided into categories using metrics such as vol- study, CNNs are built for the elastic localization in 3-D micro-
ume fraction and pair-correlation functions so that these different structures of high contrast composite microstructures using the
categories would have low inter-similarity and high intra- typical CNN architecture illustrated in Fig. 2.
similarity. In the second step, the reduced set of features in
Ref. [19] were used to characterize the microstructures in each
category and the random forest algorithms were extended sepa- 3.1. Convolutional layer
rately to these categories to predict microscale elastic strain field.
The multi-agent feature extraction approach did result in a trained Convolutional layer is the core component of CNN and it uses
model that exhibited higher accuracy compared to the single-agent filters to extract the salient features from images. The readily
feature extraction methods. accessible implementations of CNN employ 2-D filters on 2-D im-
ages. Since our application involves 3-D microstructures, we need
to find a suitable workaround for this limitation that does not
2.2. First-order localization series of MKS framework
significantly increase the computational cost involved in the
training of the model. In this study, we accomplished this by
Another data-driven method used as benchmark in this paper is
treating the 3-D neighborhood around the focal voxel as multiple
the localization series of MKS framework. MKS framework ad-
channels of 2-D layers, where each layer is then convolved with a 2-
dresses the localization problem in the form of a series expansion
D CNN filter. While this approach provides us a pseudo 3-D filter, it
derived from the statistical continuum theories [12,13]. Each term
should be recognized that it is not exactly the same as using 3-D
in the series is defined as a convolution product of the Green's
filters. For example, the trained model in the approach used here
function based kernel over the microstructure descriptor. For the
is likely to be sensitive to which planes are selected for converting
elastic localization problem, the first-order MKS expansion for a
the 3-D neighborhoods into stacks of 2-D layers. In this work, the 2-
macroscale imposed strain component 〈ε〉 can be written mathe-
D layers are selected perpendicular to the maximum principal
matically as
strain direction.
!* +
XX A general convolution layer in the present implementation takes
εs ¼ ahr mhsþr ε (1) as input N channels of 2-D images, where each image (i.e., each
r h channel) is defined on a uniform square grid of voxels. The N
channels of 2-D images are then convolved with multiple sets of N
where mhs provides the digital representation of the microstructure 2-D CNN filters, each of size m  m. The common practice is to use
(reflects the volume fraction of material local state h occupying the an odd value for m (ensures symmetry in the number of voxels on
spatial bin or voxel s [51]), r systematically indexes all of the either side of the central voxel in the convolution). The convolution
neighboring voxels near the focal voxel of interest, ahr denote the employed in this process can be mathematically expressed as
model fit parameters (or weights) that need to be trained using the
available data, and εs denotes the localized strain (model output) in X m1
N1 X X  
m1 m1 m1
the focal voxel of interest. In the case where each spatial voxel is Oðx; yÞ ¼ i¼0 I x þ i; y  þ j; k
2 2
occupied by one distinct phase, mhs takes the value of either 0 or 1 k¼0 j¼0
[22,52e55]. In the MKS framework, ahr are called influence co-  wði; j; kÞ þ b
efficients and are usually calibrated to the data aggregated from
(2)
executing finite element simulations on a large number of digitally
created microstructures. The model calibration is often pursued where I and O denote the input and output, x and y index the po-
using standard linear regression techniques. There are two major sition of voxels in an input 2-D image, and w and b denote the
distinctive features of the model form presented in Eq. (1): 1) The corresponding weights and bias captured by a set of filters (to be
influence coefficients are defined only in terms of relative distance r trained). It should be noted that each set of 2-D filters produces a 2-
to the focal voxel of interest, s. Hence, they are independent of D output image (by using Eq. (2) for all x and y values in an input 2-
microstructure morphology. Once they are obtained, they can be D image). Each 2-D output image from the convolution then serves
used to predict the response field of any new microstructure of the as an independent input channel for the next convolution layer.
same material system. 2) Since the expression shown in Eq. (1) In CNN approaches, the output O from Eq. (2) is passed through
involves a convolution product, fast Fourier transform (FFT) algo- an activation function to obtain the final output at voxel ðx; yÞ from
rithms can be exploited for computational efficiency. Consequently, the convolution layer (i.e., the use of an activation function is an
the MKS predictions for new microstructures can be obtained inherent component of the convolution layer). The main purpose of
several orders of magnitude faster compared to the physics-based the activation function is to capture the nonlinear complex func-
finite element approaches. tional mappings between input and output (note that the convo-
lution operation described above involved only a linear mapping).
3. Convolutional neural networks for localization One of the most commonly used activation function in deep
learning is ReLU (rectified linear unit) [63] formulated as f ðxÞ ¼
Convolutional neural networks (CNN), introduced by LeCun maxð0;xÞ. This activation function is applied to all voxels in the 2-D
et al. [56], are widely applied in computer vision field to solve output image from the convolution, and the output image is then
338 Z. Yang et al. / Acta Materialia 166 (2019) 335e345

Fig. 2. Illustration of a typical convolutional neural network (CNN) used in this study for building localization linkages.

called a feature map. Thus multiple feature maps (one per channel !
per convolutional layer) produced in this manner are expected to X
n1
o¼f xðiÞ  wðiÞ þ b (4)
capture the diverse range of salient features in the input that are
i¼0
strongly correlated to the selected output. Note that if one desires
to keep the size of the input images (describing the neighborhood where o represents the output of this neuron, xðiÞ denotes the ith
of a focal voxel) and of the output feature maps of the convolutional input, n is the number of inputs to the fully connected layer, and f ð:Þ
layer as same, zero padding needs to be applied when considering represents the activation function. As in convolutional layer, ReLU is
focal voxels in the boundary of the input microstructure. The the most commonly used activation function in the fully connected
alternative to zero-padding is to allow the shrinking of the output layer, except that in the final output layer, a linear activation
feature maps by ignoring focal voxels in the boundary regions of function is used for regression problems and softmax activation
the given representative volume element of the microstructure. function is used for classification problems.

3.2. Pooling layer 3.4. Proposed deep learning model

Pooling layer is used to reduce the dimensionality, while It is logical that the details of the local neighborhood strongly
retaining the most important information in the data. Average influence the accuracy of predictions from localization models. This
pooling is one of the most widely used pooling operations, and neighborhood information is usually referred as the higher-order
essentially amounts to coarsening of the image. Mathematically, microstructure information [51,64]. It is also well-known that
one can express this as more higher-order information (i.e., more details of the neighbor-
hood) are needed for localization models in high contrast com-
X
s1 X
s1 posite material systems. In Refs. [19,20], a 5  5  5 cube centered
Iðs  x þ i; s  y þ jÞ
Oðx; yÞ ¼ (3) on the focal voxel was used to build the localization models for a
s2
j¼0 i¼0 composite with a contrast of 10 in the Young's moduli of its con-
stituents. In order to predict microscale elastic strain field of higher
where s denotes the coarsening length scale (in number of voxels), contrast composites studied here, a 11  11  11 cube centered on
and ðx; yÞ identifies a voxel in the output of the feature map after the focal voxel was taken as input to the deep learning model. As
applying the average pooling layer. The pooling layer is applied for already mentioned, this means that the input is treated as 11
each feature map, which means the number of input and output channels of 2-D images of size 11  11  11 voxels. The input im-
feature maps is the same. In most cases, the pooling layer results in ages had only binary values, i.e., 0s and 1s. The input images were
a reduction of the number of the spatial voxels by an integer factor. transformed such that the values at each voxel were assigned
The most important advantage of the pooling operation is that it as 0.5 or 0.5 (replacing 0 and 1, respectively) as this was found to
reduces the computational requirements in training the mathe- improve the performance of the convolution layers. In addition, the
matical model (i.e. regression, classification, etc.) as it reduces the local strains were scaled by a factor of 10000 to avoid numerical
number of dimensions flowing from convolutional layers to the issues associated with their small magnitudes (recall that these are
fully connected layers. This also helps in the case of regression elastic strains).
problems where overfitting is a major problem, especially for cases Different CNNs and CNN architectures were explored to find the
involving a large number of features. best trained model for the present application. The architecture of
deep learning model is crucial to determine its learning capability.
Thus different architectures of CNNs consisting of convolutional
3.3. Fully connected layer layers with zero-padding, pooling layer and fully connected layers,
were first explored. More specifically, the depth of CNN was grad-
In the final stage of CNN, the output of several stacked con- ually increased from 7 layers to 18 layers by stacking additional
volutional layers and/or pooling layers are flattened to a one convolutional layers and pooling layers. However, the accuracy of
dimensional vector, and this vector is fed into a fully connected these trained models was inadequate, and did not improve much
layer. A fully connected layer is the same as a layer of conventional with these changes. There might be two reasons for this relatively
artificial neural network, which consists of multiple neurons. The poor performance: (1) The size of input image is relatively small
operation of a neuron can be formulated as compared with images in conventional computer vision field.
 Z. Yang et al. / Acta Materialia 166 (2019) 335e345 339

Consequently, reducing the image size by using pooling layer might (i.e., the total predicted strain on each MVE is 0.001). The datasets
lose too much microstructural information. (2) Due to fact that and the performance of the proposed deep learning models are
input images are binary and small, all the voxels in the images could presented in the next section.
have significant effect on the final predictions. Thus adding 0s
around the boundary by using zero-padding might confuse the 4. Results and discussion
training process. Therefore, we decided to use a six-layer CNN that
consisted of convolutional layers without zero-padding and fully 4.1. Datasets
connected layers. Because convolutional layer is the core compo-
nent of CNN and it uses filters to capture the salient features from As mentioned before, the goal of this paper is to predict micro-
the images, the number of filters in each convolutional layer was scale elastic strain field for 3-D high contrast elastic composites
systematically varied. More specifically, the number of filters in using CNN approaches. In this study, we used digitally generated
each convolutional layer was gradually increased to find the best microstructures and finite element (FE) tools to produce the data
number of filters in each convolutional layer for the present needed to train the linkages. Even though using 3-D microstructures
application. After that, the number of neurons in fully connected obtained via experimental methods sounds more attractive, such
layers was systematically varied so that the salient features datasets are scarce. Since deep learning approaches require more
captured by the conovlutional layers could be effectively utilized by datasets than traditional machine learning techniques, accumu-
fully connected layers to make accurate predictions. Finally, L2 lating a training set from experiments is expensive and time-
regularization with different penalty strengths were explored to consuming. Hence, simulation datasets are employed in this study
avoid overfitting. to explore the viability of the deep learning techniques for predicting
The architecture of the best CNN model resulting from the many microscale elastic strain field for 3-D microstructures.
trials conducted for the present application is shown in Table 1. In To train CNN for localization, two ensembles of microstructures
this model, convolutional layers 1 and 2 have 128 and 256 3  3 with contrasts of 10 and 50 were generated. Each 3-D volume of
filters, respectively, and zero-padding is not employed in the microstructure is referred to as microscale volume element (MVE)
convolution operation. The convolutional layers are followed by and they were discretized into a uniform grid of size 21  21  21.
two fully connected layers where the first and second layers Each voxel is occupied by one of the phases. The soft and hard
contain 2048 and 1024 neurons, respectively. All the weights in the phase are indicated by 0 and 1, respectively. The phases were
CNN are initialized by normalized initialization [65]. This initiali- assigned to the voxels in a randomized manner. The contrast 10
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
zation method samples a U½r; r with r ¼ fan inþfan 6 where ensemble have MVEs with volume fractions ranging from 0.01% to
out
99.43% for hard phase. On the other hand, the 50 contrast ensemble
fan in and fan out are the number of inputs and outputs of the had volume fractions in the range of 0.01%e99.99% for hard phase.
layer, respectively. ReLU is used as activation function for all the Once the MVEs were generated, finite element (FE) simulations
convolutional layers and fully connected layers except the output were executed by using the commercial software of ABAQUS [67].
layer where linear activation function is used. In order to avoid Each voxel in the MVE was converted into a 3-D 8-noded C3D8
overfitting, L2 regularization with 0.0001 penalty strength is element [67]. All simulations employed periodic boundary condi-
applied in each convolutional layer and fully connected layer. tions where all macroscale strain values had zero values except the
During the training, batch size was set as 252 data points (i.e., normal strain in direction 1 (i.e., hεij i ¼ 0 except hε11 is0 where i; j ¼
252 11  11  11 cubes centered at corresponding focal voxels 1; 2; 3 and 〈〉 represent the volume average). The macroscale normal
were propagated through the network in each training iteration), strain component in direction 1 is taken as hε11 i ¼ 0:001 for both
and mean squared error was used as loss function. Adam optimizer ensembles. The local response field of interest was selected as ε11 . It
[66] was used as it is a more effective optimization algorithm than should be noted that the strategy proposed in this paper can be
conventional stochastic gradient descent algorithm. For the repeated for a total of six macroscale strain conditions to obtain the
parameter settings of Adam, the learning rate was 0.001, b1 value full field of deformation using the superposition principle [68,69].
was 0.9 and b2 value was 0.999. Early stopping technique was As mentioned before, two ensembles of datasets with contrasts
applied to determine when to terminate the training process. More of 10 and 50 were generated to evaluate the viability of deep
specifically, the training process was terminated when the value of learning approaches developed for elastic localization linkages.
loss function on validation set did not improve over 10 epochs (i.e., Both microscale constituents (i.e., distinct phases) were assumed to
10 complete passes through the entire dataset). be isotropically elastic and the Poisson ratio, n, for both phases in
There is a post-processing step after the training is terminated both ensembles was taken as 0.3. The contrast in elastic composites
and the predictions are computed. Since the total strain on a MVE of was defined as the ratio between the Young's modulus of the two
FE model is fixed (i.e., 0.001 for both contrast 10 and 50 composites distinct phases. For contrast-10 dataset, the Young's modulus of
materials datasets discussed next), the predicted local strain of hard and soft phases were selected as E1 ¼ 120GPa and E2 ¼
focal voxel is multiplied by a scaling factor 0:001=e where e is the 12GPa, respectively. On the other hand, for contrast-50 dataset, the
total predicted strain on the MVE so that the total strain on each modulus of the hard phase is kept the same while the soft phase
MVE of ground truth (i.e., FE model) and predictions are the same was assigned a Young's modulus of 2:4GPa. An example MVE with
contrast-10 and its elastic strain field of ε11 is depicted in Fig. 3.
Table 1 For the ensemble of contrast-10, 2500 MVEs were generated.
The dimensionality of each layers in the proposed CNN (bs. Denotes the batch 1200 and 300 of these MVEs were used for training and validation
size). of CNN, respectively. The remaining of 1000 MVEs were not
Layer Dimension included in the training and used only as a testing set. To be more
Input Layer bs:  11  11  11
specific, there were 100 vol fraction categories and 25 MVEs in each
Convolutional Layer 1 bs:  9  9  128 volume fraction category for this dataset. For each volume fraction,
Convolutional Layer 2 bs:  7  7  256 12 MVEs were randomly selected for training set, 3 MVEs were
Fully Connected Layer 1 bs:  2048 randomly selected for validation set and rest 10 MVEs were used for
Fully Connected Layer 2 bs:  1024 testing set. In this way, all the three sets contained representations
Output Layer bs:  1
from all the volume fraction. Each MVE has 21  21  21ð¼ 9261Þ
340 Z. Yang et al. / Acta Materialia 166 (2019) 335e345

Fig. 3. Visualization of an example contrast-10 MVE (left) and its strain field (right). The white and black voxels in the MVE correspond to hard and soft phases, respectively.

voxels where each voxel provides one data point for building and benchmark methods. The proposed deep learning model achieves
evaluating the desired models. Hence, training, validation and test 3.07% average MASE for the entire testing set, and the standard
sets had 1200  9261ð ¼ 11; 113; 200Þ, 300  9261ð ¼ 2; 778; 300Þ, deviation of MASE across all the MVEs in the testing set is 1.22%. By
and 1000  9261ð¼ 9; 261; 000Þ voxels, respectively. On the other comparing with benchmark methods, we observe that the pro-
hand, 3000 MVEs were generated for contrast-50. Using the similar posed deep learning model improves the previous best prediction
data splitting strategy, this ensemble was split into 2000 training, performance in terms of MASE by as much as around 61.8% (1e3.07/
500 validation and 500 testing MVEs, respectively. This resulted in 8.04), which shows that deep learning model can produce a more
2000  9261ð ¼ 18; 522; 000Þ, 300  9261ð ¼ 2; 778; 300Þ, and accurate predictive model. Further, to evaluate and quantify the
500  9261ð¼ 4; 630; 500Þ data points for training, validation, and robustness of the proposed deep learning model, we constructed
testing of the ensemble of contrast-50. Both datasets are large and evaluated 10 deep learning models on different training-
enough to develop deep learning models. testing splits (but same hyper-parameters). The average MASE
We used Python 2.7 and Keras [70], which is a high-level neural across the 10 trials is 3.23% with a standard deviation of 0.22%,
networks library built on top of TensorFlow [71] to implement deep which is relatively small, suggesting that the proposed deep
learning model. Scikit-learn [72] and PyMKS [73] are used for learning model is quite robust. In addition, in Ref. [20], training data
implementing benchmark methods. CNN trials were carried out on is divided into several groups by some dividing criterion and then
a NVIDIA DIGITS DevBox with 4 TITAN X GPUs with 12 GB memory individual model is developed for each data group. However, the
for each GPU and Core i7-5930 K 6 Core 3.5 GHz CPU with 64 GB proposed deep learning model is directly trained on overall training
DDR4 RAM. In order to get a fair comparison, model performance is set, which produces a model with better generalization.
evaluated by the mean absolute strain error (MASE) [19], which is Fig. 4 shows the plot of MASE for each individual MVE in testing
defined as below, set versus volume fraction with error bars of contrast-10 compos-
ites dataset using the proposed deep learning model. We can
S  
1X  p s 
ps  b observe that the MASE for MVEs with around 40% volume fraction
e¼  100% (5)
S s¼1  hε11 i  for hard phase has the largest error and variance (i.e. large average
MASE and standard deviation).
where hε11 i is the macroscopic strain tensor component applied to The results in Fig. 4 assess the performance of deep learning
MVE via periodic boundary conditions. This metric is used to model (DL) on an average basis. Since the problem at hand is elastic
measure average error for a single MVE. ps and bp s denote the local localization, we also compare the local strain fields obtained from
strain in the voxel s from the FE model and CNN model, respec- FE method to DL and MKS methods. The comparison of FE, DL and
tively. The performance of predicting microscale elastic strain field MKS methods at the level of individual voxel is shown in Fig. 5 for
associated with CNN is also compared to benchmark methods using two slices selected to represent results from the best and worst
the same error metric. predictions (based on MASE values of DL). For the MVE with best
MASE, the strain field histogram of DL predictions matches that of
ground truth very well, except that the curve of Phase 1 prediction
4.2. Results of contrast-10 composites dataset shifts slightly to the right. The difference of the strain map of a
randomly selected slide between DL predictions and ground truth
Table 2 shows the comparison of the average MASE of entire is indistinguishable, while the performance of MKS method is
testing set for the proposed deep learning model and the significant worse. On the other hand, even for the MVE with worst
MASE, the trend of strain field histogram of DL predictions matches
that of ground truth very well. The differences of DL prediction and
Table 2 ground truth in the strain field maps are unnoticeable, while MKS
Results comparison of different models for contrast-10 dataset (standard deviations
are available only for MKS method and deep learning model because studies
method tends to underestimate the strain field.
describing other benchmarks did not present results of standard deviation).

Method Average MASE for testing set 4.3. Results of contrast-50 composite dataset
Single-agent based method [19] 13.02%
MKS method 10:86%±4:30% The same architecture and hyperparameters used to train the
Multi-agent based method [20] 8.04% ensemble of contrast-10 were also used to train the CNN on the
Deep learning model 3:07%±1:22%
ensemble of contrast-50. For this case study, it was only possible to
 Z. Yang et al. / Acta Materialia 166 (2019) 335e345 341

Table 3
Results comparison of different models for contrast-50 dataset.

Method Average MASE for testing set

MKS method 26:46%±6:91%

Deep learning model 5:71%±2:46%

constructed and evaluated 10 deep learning models on different

training-testing splits (but same hyper-parameters). The average
MASE across the 10 trials is 5.38% with a standard deviation of
0.22%, which is again relatively small, suggesting that the proposed
deep learning model is quite robust for contrast-50 dataset as well.
The superior performance of the CNN approach over the MKS
method indicates tremendous promise for a feature-engineering
free approach to predict microscale elastic strain field. This is of
tremendous value to multiscale materials design efforts in higher
contrast composites as the identification of the salient features
has been the central hurdle. Furthermore, the fact that the same
CNN architecture has provided excellent predictive models for
Fig. 4. Plot of MASE for each individual MVE in testing set versus volume fraction with both contrast-10 and contrast-50 datasets supports the conclusion
error bars of contrast-10 dataset using the proposed deep learning model. that the use of deep learning approaches offers a higher
generalization.
Fig. 6 shows the plot of MASE for each individual MVE in testing
use one benchmark as the studies describing other benchmarks did
set versus volume fraction with error bars of contrast-50 compos-
not present results on contrast-50 dataset, nor is their code avail-
ites dataset using the proposed deep learning model. The MASE for
able. The results in Table 3 show the comparison of average MASE of
MVEs with around 23% volume fraction for hard phase has the
entire testing set between proposed deep learning model and the
largest error and variance (i.e. large average MASE and standard
first-order MKS method. The proposed deep learning model ach-
deviation), while the MVEs with around 80% volume fraction for
ieves 5.71% average MASE on entire testing set, and the standard
hard phase has the best predictions (i.e. small average MASE and
deviation of MASE across all the MVEs in the testing set is 2.46%. In
standard deviation).
contrast, the MASE of the benchmark method is 26.46%, which is
To visually demonstrate the predictive power of deep learning
significantly higher. In other words, the proposed deep learning
model (DL) for contrast-50 dataset, the MVEs with best MASE and
model improves the performance by as much as around 78.4%
worst MASE are selected. Fig. 7 shows the comparison of FE, DL and
(1e5.71/26.46) in terms of MASE. Further, to evaluate and quantify
MKS method at the level of individual voxels. The comparisons are
the robustness of the proposed deep learning model, we

Fig. 5. Comparison of FE model (i.e. ground truth), deep learning (DL) method and MKS method predictions for contrast-10 dataset. (a) Strain fields histogram of the MVE with the
best MASE for FE and DL. (b) Strain field map of a randomly selected slide in the MVE with the best MASE for FE, DL and MKS. (c) Strain fields histogram of the MVE with the worst
MASE for FE and DL. (d) Strain field map of a randomly selected slide in the MVE with the worst MASE for FE, DL and MKS.
342 Z. Yang et al. / Acta Materialia 166 (2019) 335e345

4.4. Interpretation of what the deep learning model learns

Though deep learning model has a striking learning capability

that is superior to the traditional machine learning methods, it
usually works like a black box due to its complex architecture and
millions of parameters. Thus, the interpretation of what the deep
learning model actually learns merits attention. In this section, we
adjust the inputs of the model and interpret what the deep learning
model has learned by analyzing their corresponding predictions.
In Ref. [19], Liu et al. identify groups of neighboring voxels based
on their distance from the focal voxel of interest. In this classifi-
cation scheme, L level neighbors include all the voxels that are at a
center-to-center Euclidean distance of square root of L from the
voxel of interest (treating each cubic voxel to be of unit length).
Fig. 8 illustrates examples of the first three level neighbors.
Empirically, lower level neighbors have larger effect on the
response of the focal voxel. In order to see if the proposed deep
learning model can capture this information, we adjust the input
data by setting the values of different level neighbors as 0. Note that
Fig. 6. Plot of MASE for each individual MVE in testing set versus volume fraction with the inputs have been rescaled from ½0; 1 to ½0:5; 0:5 during the
error bars of contrast-50 dataset using the proposed deep learning model. training, thus setting the value as 0 eliminates the contribution
from that voxel. Then we can predict the modified response, and
assess the percentage contribution from the different levels of
made for a randomly selected slice from 3-D MVEs that exhibited neighbors. For this purpose, we define a residual percentage E
the best and the worst MASE of DL. For the MVE with the best
MASE, the trend and the peak of the strain field histogram of DL
jb
y  yj
predictions and ground truth accurately match each other. The E¼  100% (6)
difference between FE and DL is almost indistinguishable from each y
other, while the MKS method tends to overestimate the strain field.
where by and y are prediction and ground truth, respectively. A total
For the MVE with the worst MASE, the overall trend of strain field
of 11 MVEs from the contrast-10 testing set were selected with a
histogram of DL predictions and ground truth match each other
difference of volume fraction around 10% (i.e., their volume frac-
very well. Even the strain maps are quite similar to each other
tions are around 1%, 9%, 19%, 29%, …,99%). Similarly, 11 MVEs from
demonstrating the fidelity of the DL models for high contrast
the contrast-50 testing set were also selected with a difference of
composites, while the MKS method underestimates the strain filed
volume fraction around 6% (i.e., their volume fractions are around
resulting in a significant error.
20%, 26%, 32%, 38%, …, 80%). For each voxel in all these MVEs, we

Fig. 7. Comparison of FE model (i.e. ground truth), deep learning (DL) method and MKS method predictions for contrast-50 dataset. (a) Strain histograms of the MVE with the best
MASE for FE and DL. (b) Strain field map of a randomly selected slide in the MVE with the best MASE for FE, DL and MKS. (c) Strain histograms of the MVE with the worst MASE for
FE and DL. (d) Strain field map of a randomly selected slide in the MVE with the worst MASE for FE, DL and MKS.
 Z. Yang et al. / Acta Materialia 166 (2019) 335e345 343

Fig. 8. Illustration of different level neighbors. (a) first, (b) second, and (c) third level neighbor.

compute the residual percentage E of original data input, then which attests to their high learning capabilities when there is a
repeated the computations by removing first, second, third, and up sufficiently large dataset. 2) Generalization: in order to get better
to twenty first neighbor level, one at a time. accuracy, multi-agent learning strategy is usually used in tradi-
Fig. 9 shows the plots of residual percentage after removing tional machine learning method. Because the variance of the data is
different level neighbors for both contrast-10 and contrast-50 decreased after it is divided into multiple data clusters, such pro-
datasets. The plots show that the residual percentage generally cessing could deteriorate the generalization of the model. In addi-
decreases with increasing level of neighbors, and the first and tion, complex problem-specific features could also make the model
second level neighbors have the most significant influence on the difficult to generalize to other material systems. In contrast, deep
response of focal voxel of interest. Since the CNN was not explicitly learning has the capability to directly train on entire raw dataset
provided this knowledge in any manner, it is indeed remarkable without feature engineering and extract enough useful micro-
that the CNN model automatically reflects this knowledge. Note structure information, which results in a model with better
that the CNN model is able to capture this knowledge at both generalization. As we know, data is usually limited in material
contrast levels. It is also remarkable that the CNN model captured science research, which hinders the applications of deep learning in
higher levels of interactions at a given L in the higher contrast the field. Thus, transfer learning [74] can provide a promising
composite compared to the lower contrast composite. Once again, alternative. Because the 3-D microstructures used in this work do
the CNN model learned this knowledge also by itself. In addition, not have any special constraints, such as volume fraction, orienta-
the curves in the plots are not decreasing monotonically. This tion, and shape of either soft or hard phase, the features learned by
means that the distance from focal voxel of interest is not the only the proposed model are general enough to well characterize gen-
factor that decides the significance of neighbors. In other words, the eral two-phase microstructures. Thus, the proposed model can
contributions from the higher level neighbors could be more sig- serve as a pre-trained model for transfer learning to assist re-
nificant than that from lower level neighbors due to their different searchers in developing models to predict materials properties of 3-
directional relation to the focal voxel of interest. The trained CNN D two-phase microstructures. Therefore, it can accelerate the ap-
model is thus found to have implicitly learned all of this salient plications of deep learning in material science research. In addition,
information to make accurate predictions. considering the advantages mentioned above, the other two main
The CNN models presented in this work show their superiority use cases for the proposed data-driven model could be, 1) It can
compared to traditional machine learning approaches mainly in allow rapid evaluation of a large number of microstructures in
two aspects. 1) Accuracy: the results show that the deep learning designing new materials meeting desired mechanical responses.
approaches can improve the model's performance by more than This is because the data-driven models are computationally cheap
50% compared with traditional machine learning approaches, compared to the numerical tools such as FEM models. In fact,

Fig. 9. Plots showing contributions of the different level neighbors in the trained CNN models. The dashed lines in both figures show the average percentage contributions of
selected MVE in the datasets, and the solid lines in both figures show the average residual percentage of a collection of selected MVEs in the datasets. (a) plot for contrast-10 dataset.
(b) plot for contrast-50 dataset.
344 Z. Yang et al. / Acta Materialia 166 (2019) 335e345

design of material microstructure using FEM models is prohibi- materials: self-consistent and mori-tanaka schemes, Int. J. Plast. 25 (6) (2009)
1024e1048.
tively expensive and its viability has not been demonstrated in any , R.A. Lebensohn, Self consistent homogenization methods for
[9] C.N. Tome
prior literature. However, the data-driven models have the poten- texture and anisotropy, in: Continuum Scale Simulation of Engineering Ma-
tial to address this gap [75e77]. 2) The data-driven models can also terials: Fundamentalsemicrostructureseprocess Applications, 2004,
help in objective calibration of the many parameters present in the pp. 473e499.
[10] V. Levin, F. Sabina, J. Bravo-Castillero, R. Guinovart-Diaz, R. Rodriguez-Ramos,
physics-based microscale models to the often limited amount of O. Valdiviezo-Mijangos, Analysis of effective properties of electroelastic
experimental data [78e80]. However, there are also limitations for composites using the self-consistent and asymptotic homogenization
this work. Because large amount of experimental data is not methods, Int. J. Eng. Sci. 46 (8) (2008) 818e834.
[11] H. Garmestani, S. Lin, B. Adams, S. Ahzi, Statistical continuum theory for large
available, the proposed model is trained only on simulation data- plastic deformation of polycrystalline materials, J. Mech. Phys. Solid. 49 (3)
sets. The effectiveness of the proposed model still needs to be (2001) 589e607.
validated with suitable experimental data. If the proposed model [12] E. Kro€ner, Bounds for effective elastic moduli of disordered materials, J. Mech.
Phys. Solid. 25 (2) (1977) 137e155.
could also be trained and validated with experimental data, it can [13] E. Kro€ner, Statistical modelling, in: Modelling Small Deformations of Poly-
offer a valuable, low cost, and accurate predictive tool for multiscale crystals, Springer, 1986, pp. 229e291.
material design efforts. [14] H. Moulinec, P. Suquet, A numerical method for computing the overall
response of nonlinear composites with complex microstructure, Comput.
Methods Appl. Mech. Eng. 157 (1e2) (1998) 69e94.
5. Conclusions [15] A. Prakash, R. Lebensohn, Simulation of micromechanical behavior of poly-
crystals: finite elements versus fast fourier transforms, Model. Simulat. Mater.
Sci. Eng. 17 (6) (2009), 064010.
In this paper, a convolutional neural network is developed and [16] R.A. Lebensohn, A.K. Kanjarla, P. Eisenlohr, An elasto-viscoplastic formulation
tested on both contrast-10 and contrast-50 composite datasets. The based on fast fourier transforms for the prediction of micromechanical fields
results show that the proposed deep learning model significantly in polycrystalline materials, Int. J. Plast. 32 (2012) 59e69.
[17] Y.C. Yabansu, D.K. Patel, S.R. Kalidindi, Calibrated localization relationships for
outperforms benchmark methods by 61.8% and 78.4% in terms of elastic response of polycrystalline aggregates, Acta Mater. 81 (2014) 151e160.
average MASE of testing set on contrast-10 and contrast-50 data- [18] Y.C. Yabansu, S.R. Kalidindi, Representation and calibration of elastic locali-
sets, respectively. It is clear from these trials that deep learning is a zation kernels for a broad class of cubic polycrystals, Acta Mater. 94 (2015)
26e35.
promising technique to build a feature-engineering-free, high ac- [19] R. Liu, Y.C. Yabansu, A. Agrawal, S.R. Kalidindi, A.N. Choudhary, Machine
curacy, low computational cost, and high generalization model to learning approaches for elastic localization linkages in high-contrast com-
study PSP linkages in complex materials systems. Moreover, since posite materials, Integrating Mater. Manuf. Innov. 4 (1) (2015) 13.
[20] R. Liu, Y.C. Yabansu, Z. Yang, A.N. Choudhary, S.R. Kalidindi, A. Agrawal,
the proposed deep learning architecture worked well for a large Context aware machine learning approaches for modeling elastic localization
range of contrasts, it can be used as a pre-trained model with in three-dimensional composite microstructures, Integrating Mater. Manuf.
transfer learning approach to establish PSP linkages for other high Innov. (2017) 1e12.
[21] G. Landi, S.R. Kalidindi, Thermo-elastic localization relationships for multi-
contrast composites materials when the dataset is limited. There
phase composites, Comput. Mater. Continua (CMC) 16 (3) (2010) 273e293.
are several possible directions for future works based on this study. [22] G. Landi, S.R. Niezgoda, S.R. Kalidindi, Multi-scale modeling of elastic response
As an example, we only show the relative importance of neighbors of three-dimensional voxel-based microstructure datasets using novel dft-
based on the distance from focal voxel in section 4.4, one could based knowledge systems, Acta Mater. 58 (7) (2010) 2716e2725.
[23] T. Fast, S.R. Kalidindi, Formulation and calibration of higher-order elastic
further investigate the relative importance of neighbors based on localization relationships using the mks approach, Acta Mater. 59 (11) (2011)
other criterion, such as the shape of phase, which might provide 4595e4605.
more insights about relationship between focal voxel and its [24] M.W. Priddy, N.H. Paulson, S.R. Kalidindi, D.L. McDowell, Strategies for rapid
parametric assessment of microstructure-sensitive fatigue for hcp poly-
neighbors. Such insights if established reliably could lead to a new crystals, Int. J. Fatig. 104 (2017) 231e242.
easy way to implement design rules for microstructures (aimed at [25] C. Przybyla, R. Prasannavenkatesan, N. Salajegheh, D.L. McDowell, Micro-
meeting a desired property). structure-sensitive modeling of high cycle fatigue, Int. J. Fatig. 32 (3) (2010)
512e525.
[26] V. Shankar, M. Valsan, K.B.S. Rao, R. Kannan, S. Mannan, S. Pathak, Low cycle
Acknowledgements fatigue behavior and microstructural evolution of modified 9cre1mo ferritic
steel, Mater. Sci. Eng., A 437 (2) (2006) 413e422.
[27] Y. Djebara, A. El Moumen, T. Kanit, S. Madani, A. Imad, Modeling of the effect
This work is supported in part by the following grants: AFOSR of particles size, particles distribution and particles number on mechanical
award FA9550-12-1-0458; NIST awards 70NANB14H012 and properties of polymer-clay nano-composites: numerical homogenization
70NANB14H191; NSF award CCF-1409601; DOE awards versus experimental results, Compos. B Eng. 86 (2016) 135e142.
[28] V. Kouznetsova, M. Geers, W. Brekelmans, Multi-scale second-order compu-
DESC0007456, DESC0014330; and Northwestern Data Science tational homogenization of multi-phase materials: a nested finite element
Initiative. solution strategy, Comput. Methods Appl. Mech. Eng. 193 (48e51) (2004)
5525e5550.
[29] N.H. Paulson, M.W. Priddy, D.L. McDowell, S.R. Kalidindi, Reduced-order
References structure-property linkages for polycrystalline microstructures based on 2-
point statistics, Acta Mater. 129 (2017) 428e438.
[1] S. Groh, E. Marin, M. Horstemeyer, H. Zbib, Multiscale modeling of the plas- [30] D.J. Luscher, D.L. McDowell, C.A. Bronkhorst, A second gradient theoretical
ticity in an aluminum single crystal, Int. J. Plast. 25 (8) (2009) 1456e1473. framework for hierarchical multiscale modeling of materials, Int. J. Plast. 26
[2] V. Vaithyanathan, C. Wolverton, L. Chen, Multiscale modeling of precipitate (8) (2010) 1248e1275.
microstructure evolution, Phys. Rev. Lett. 88 (12) (2002) 125503. [31] J.E. Aarnes, S. Krogstad, K.-A. Lie, A hierarchical multiscale method for two-
[3] S. Chandra, M. Samal, V. Chavan, S. Raghunathan, Hierarchical multiscale phase flow based upon mixed finite elements and nonuniform coarse grids,
modeling of plasticity in copper: from single crystals to polycrystalline ag- Multiscale Model. Simul. 5 (2) (2006) 337e363.
gregates, Int. J. Plast. 101 (2018) 188e212. [32] R.A. Lebensohn, A.D. Rollett, P. Suquet, Fast fourier transform-based modeling
[4] J. Knap, C. Spear, K. Leiter, R. Becker, D. Powell, A computational framework for for the determination of micromechanical fields in polycrystals, JOM (J. Occup.
scale-bridging in multi-scale simulations, Int. J. Numer. Methods Eng. 108 (13) Med.) 63 (3) (2011) 13e18.
(2016) 1649e1666. [33] T. Fast, S.R. Niezgoda, S.R. Kalidindi, A new framework for computationally
[5] X. Wu, G. Proust, M. Knezevic, S. Kalidindi, Elasticeplastic property closures efficient structureestructure evolution linkages to facilitate high-fidelity scale
for hexagonal close-packed polycrystalline metals using first-order bounding bridging in multi-scale materials models, Acta Mater. 59 (2) (2011) 699e707.
theories, Acta Mater. 55 (8) (2007) 2729e2737. [34] M. Guillaumin, J. Verbeek, C. Schmid, Is that you? metric learning approaches
[6] P.P. Castan~ eda, Second-order homogenization estimates for nonlinear com- for face identification, in: Computer Vision, 2009 IEEE 12th International
posites incorporating field fluctuations: Idtheory, J. Mech. Phys. Solid. 50 (4) Conference on, IEEE, 2009, pp. 498e505.
(2002) 737e757. [35] C. Huang, S. Zhu, K. Yu, Large Scale Strongly Supervised Ensemble Metric
[7] B.L. Adams, M. Lyon, B. Henrie, Microstructures by design: linear problems in Learning, with Applications to Face Verification and Retrieval, arXiv preprint
elasticeplastic design, Int. J. Plast. 20 (8e9) (2004) 1577e1602. arXiv:1212.6094.
[8] S. Mercier, A. Molinari, Homogenization of elasticeviscoplastic heterogeneous [36] A. Mignon, F. Jurie, Pcca: a new approach for distance learning from sparse
 Z. Yang et al. / Acta Materialia 166 (2019) 335e345 345

pairwise constraints, in: Computer Vision and Pattern Recognition (CVPR), Process. 14 (9) (2005) 1360e1371.
2012 IEEE Conference on, IEEE, 2012, pp. 2666e2672. [59] P. Sermanet, Y. LeCun, Traffic sign recognition with multi-scale convolutional
[37] B. Ma, X. Ban, H. Huang, Y. Chen, W. Liu, Y. Zhi, Deep learning-based image networks, in: Neural Networks (IJCNN), the 2011 International Joint Confer-
segmentation for al-la alloy microscopic images, Symmetry 10 (4) (2018) 107. ence on, IEEE, 2011, pp. 2809e2813.
[38] M.L. Silvoster, V. Govindan, Enhanced cnn based electron microscopy image [60] D.C. Ciresan, U. Meier, J. Masci, L. Maria Gambardella, J. Schmidhuber, Flexible,
segmentation, Cybern. Inf. Technol. 12 (2) (2012) 84e97. high performance convolutional neural networks for image classification, in:
[39] M. Ziatdinov, A. Maksov, S.V. Kalinin, Learning surface molecular structures IJCAI Proceedings-international Joint Conference on Artificial Intelligence, vol.
via machine vision, npj Computational Materials 3 (1) (2017) 31. 22, Barcelona, Spain, 2011, p. 1237.
[40] R. Liu, A. Agrawal, W.-k. Liao, A. Choudhary, M. De Graef, Materials discovery: [61] T. Wang, D.J. Wu, A. Coates, A.Y. Ng, End-to-end text recognition with con-
understanding polycrystals from large-scale electron patterns, in: Big Data volutional neural networks, in: Pattern Recognition (ICPR), 2012 21st Inter-
(Big Data), 2016 IEEE International Conference on, IEEE, 2016, pp. 2261e2269. national Conference on, IEEE, 2012, pp. 3304e3308.
[41] R. Cang, H. Li, H. Yao, Y. Jiao, Y. Ren, Improving Direct Physical Properties [62] S. Ji, W. Xu, M. Yang, K. Yu, 3d convolutional neural networks for human
Prediction of Heterogeneous Materials from Imaging Data via Convolutional action recognition, IEEE Trans. Pattern Anal. Mach. Intell. 35 (1) (2013)
Neural Network and a Morphology-aware Generative Model, arXiv preprint 221e231.
arXiv:1712.03811. [63] V. Nair, G.E. Hinton, Rectified linear units improve restricted Boltzmann ma-
[42] R. Liu, L. Ward, C. Wolverton, A. Agrawal, W. Liao, A. Choudhary, Deep learning chines, in: Proceedings of the 27th International Conference on Machine
for chemical compound stability prediction, in: Proceedings of ACM SIGKDD Learning, ICML-10), 2010, pp. 807e814.
Workshop on Large-scale Deep Learning for Data Mining, DL-KDD), 2016, [64] B.L. Adams, S. Kalidindi, D.T. Fullwood, Microstructure-sensitive Design for
pp. 1e7. Performance Optimization, Butterworth-Heinemann, 2013.
[43] Z. Yang, Y.C. Yabansu, R. Al-Bahrani, W.-k. Liao, A.N. Choudhary, S.R. Kalidindi, [65] X. Glorot, Y. Bengio, Understanding the difficulty of training deep feedforward
A. Agrawal, Deep learning approaches for mining structure-property linkages neural networks, Aistats 9 (2010) 249e256.
in high contrast composites from simulation datasets, Comput. Mater. Sci. 151 [66] D. Kingma, J. Ba, Adam: A Method for Stochastic Optimization, arXiv preprint
(2018) 278e287. arXiv:1412.6980.
[44] A. Cecen, H. Dai, Y.C. Yabansu, L. Song, S.R. Kalidindi, Material structure- [67] Karlsson Hibbett, Sorensen, ABAQUS/standard: User's Manual, vol. 1, Hibbitt,
property linkages using three-dimensional convolutional neural networks, Karlsson & Sorensen, 1998.
Acta Mater. 146 (2018) 76e84. [68] H.F. Al-Harbi, G. Landi, S.R. Kalidindi, Multi-scale modeling of the elastic
[45] R. Kondo, S. Yamakawa, Y. Masuoka, S. Tajima, R. Asahi, Microstructure response of a structural component made from a composite material using
recognition using convolutional neural networks for prediction of ionic con- the materials knowledge system, Model. Simulat. Mater. Sci. Eng. 20 (5)
ductivity in ceramics, Acta Mater. 141 (2017) 29e38. (2012), 055001.
[46] X. Li, Y. Zhang, H. Zhao, C. Burkhart, L. C. Brinson, W. Chen, A Transfer Learning [69] S.R. Kalidindi, G. Landi, D.T. Fullwood, Spectral representation of higher-order
Approach for Microstructure Reconstruction and Structure-property Pre- localization relationships for elastic behavior of polycrystalline cubic mate-
dictions, arXiv preprint arXiv:1805.02784. rials, Acta Mater. 56 (15) (2008) 3843e3853.
[47] R. Cang, Y. Xu, S. Chen, Y. Liu, Y. Jiao, M.Y. Ren, Microstructure representation [70] F. Chollet, Keras, 2015. https://github.com/fchollet/keras.
and reconstruction of heterogeneous materials via deep belief network for [71] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G.S. Corrado,
computational material design, J. Mech. Des. 139 (7) (2017), 071404. A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving,
[48] Z. Yang, X. Li, L.C. Brinson, A. Choudhary, W. Chen, A. Agrawal, Microstructural M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mane ,
materials design via deep adversarial learning methodology, J. Mech. Des. 140 R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner,
(11) (2018) 10. I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Vie gas,
[49] X. Li, Z. Yang, L.C. Brinson, A. Choudhary, A. Agrawal, W. Chen, A deep O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, X. Zheng, TensorFlow:
adversarial learning methodology for designing microstructural material Large-scale Machine Learning on Heterogeneous Systems, 2015 software
systems, in: ASME 2018 International Design Engineering Technical Confer- available from: tensorflow.org, http://tensorflow.org/.
ences and Computers and Information in Engineering Conference, American [72] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,
Society of Mechanical Engineers, 2018. V02BT03A008eV02BT03A008. M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos,
[50] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: machine
A. Courville, Y. Bengio, Generative adversarial nets, in: Advances in Neural learning in Python, J. Mach. Learn. Res. 12 (2011) 2825e2830.
Information Processing Systems, 2014, pp. 2672e2680. [73] D. Wheeler, D. Brough, T. Fast, S. Kalidindi, A. Reid, Pymks, Materials
[51] B.L. Adams, X.C. Gao, S.R. Kalidindi, Finite approximations to the second-order Knowledge System in python, figshare, 2014, 2016.
properties closure in single phase polycrystals, Acta Mater. 53 (13) (2005) [74] S.J. Pan, Q. Yang, A survey on transfer learning, IEEE Trans. Knowl. Data Eng. 22
3563e3577. (10) (2010) 1345e1359.
[52] S.R. Kalidindi, S.R. Niezgoda, G. Landi, S. Vachhani, T. Fast, A novel framework [75] R. Liu, A. Kumar, Z. Chen, A. Agrawal, V. Sundararaghavan, A. Choudhary, A
for building materials knowledge systems, Comput. Mater. Continua (CMC) 17 predictive machine learning approach for microstructure optimization and
(2) (2010) 103e125. materials design, Sci. Rep. 5.
[53] S.R. Niezgoda, A.K. Kanjarla, S.R. Kalidindi, Novel microstructure quantifica- [76] B. Meredig, A. Agrawal, S. Kirklin, J.E. Saal, J. Doak, A. Thompson, K. Zhang,
tion framework for databasing, visualization, and analysis of microstructure A. Choudhary, C. Wolverton, Combinatorial screening for new materials in
data, Integrating Mater. Manuf. Innov. 2 (1) (2013) 3. unconstrained composition space with machine learning, Phys. Rev. B 89 (9)
[54] P. Altschuh, Y.C. Yabansu, J. Ho €tzer, M. Selzer, B. Nestler, S.R. Kalidindi, Data (2014), 094104.
science approaches for microstructure quantification and feature identifica- [77] A. Agrawal, A. Choudhary, Perspective: materials informatics and big data:
tion in porous membranes, J. Membr. Sci. 540 (2017) 88e97. realization of the “fourth paradigm” of science in materials science, Apl.
[55] A. Iskakov, Y.C. Yabansu, S. Rajagopalan, A. Kapustina, S.R. Kalidindi, Appli- Mater. 4 (5) (2016), 053208.
cation of spherical indentation and the materials knowledge system frame- [78] P. Fernandez-Zelaia, V.R. Joseph, S.R. Kalidindi, S.N. Melkote, Estimating me-
work to establishing microstructure-yield strength linkages from carbon steel chanical properties from spherical indentation using bayesian approaches,
scoops excised from high-temperature exposed components, Acta Mater. 144 Mater. Des. 147 (2018) 92e105.
(2018) 758e767. [79] D.K. Patel, S.R. Kalidindi, Estimating the slip resistance from spherical nano-
[56] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to indentation and orientation measurements in polycrystalline samples of cubic
document recognition, Proc. IEEE 86 (11) (1998) 2278e2324. metals, Int. J. Plast. 92 (2017) 19e30.
[57] H. Schulz, S. Behnke, Learning object-class segmentation with convolutional [80] D.K. Patel, H.F. Al-Harbi, S.R. Kalidindi, Extracting single-crystal elastic con-
neural networks, in: ESANN, 2012. stants from polycrystalline samples using spherical nanoindentation and
[58] F. Ning, D. Delhomme, Y. LeCun, F. Piano, L. Bottou, P.E. Barbano, Toward orientation measurements, Acta Mater. 79 (2014) 108e116.
automatic phenotyping of developing embryos from videos, IEEE Trans. Image